1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/castnode.hpp" 29 #include "opto/connode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/divnode.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/matcher.hpp" 34 #include "opto/mulnode.hpp" 35 #include "opto/movenode.hpp" 36 #include "opto/opaquenode.hpp" 37 #include "opto/rootnode.hpp" 38 #include "opto/subnode.hpp" 39 40 //============================================================================= 41 //------------------------------split_thru_phi--------------------------------- 42 // Split Node 'n' through merge point if there is enough win. 43 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 44 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 45 // ConvI2L may have type information on it which is unsafe to push up 46 // so disable this for now 47 return NULL; 48 } 49 50 // Splitting range check CastIIs through a loop induction Phi can 51 // cause new Phis to be created that are left unrelated to the loop 52 // induction Phi and prevent optimizations (vectorization) 53 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check()) { 54 return NULL; 55 } 56 57 int wins = 0; 58 assert(!n->is_CFG(), ""); 59 assert(region->is_Region(), ""); 60 61 const Type* type = n->bottom_type(); 62 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 63 Node *phi; 64 if (t_oop != NULL && t_oop->is_known_instance_field()) { 65 int iid = t_oop->instance_id(); 66 int index = C->get_alias_index(t_oop); 67 int offset = t_oop->offset(); 68 phi = new PhiNode(region, type, NULL, iid, index, offset); 69 } else { 70 phi = PhiNode::make_blank(region, n); 71 } 72 uint old_unique = C->unique(); 73 for (uint i = 1; i < region->req(); i++) { 74 Node *x; 75 Node* the_clone = NULL; 76 if (region->in(i) == C->top()) { 77 x = C->top(); // Dead path? Use a dead data op 78 } else { 79 x = n->clone(); // Else clone up the data op 80 the_clone = x; // Remember for possible deletion. 81 // Alter data node to use pre-phi inputs 82 if (n->in(0) == region) 83 x->set_req( 0, region->in(i) ); 84 for (uint j = 1; j < n->req(); j++) { 85 Node *in = n->in(j); 86 if (in->is_Phi() && in->in(0) == region) 87 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 88 } 89 } 90 // Check for a 'win' on some paths 91 const Type *t = x->Value(&_igvn); 92 93 bool singleton = t->singleton(); 94 95 // A TOP singleton indicates that there are no possible values incoming 96 // along a particular edge. In most cases, this is OK, and the Phi will 97 // be eliminated later in an Ideal call. However, we can't allow this to 98 // happen if the singleton occurs on loop entry, as the elimination of 99 // the PhiNode may cause the resulting node to migrate back to a previous 100 // loop iteration. 101 if (singleton && t == Type::TOP) { 102 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 103 // irreducible loop may not be indicated by an affirmative is_Loop()); 104 // therefore, the only top we can split thru a phi is on a backedge of 105 // a loop. 106 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 107 } 108 109 if (singleton) { 110 wins++; 111 x = ((PhaseGVN&)_igvn).makecon(t); 112 } else { 113 // We now call Identity to try to simplify the cloned node. 114 // Note that some Identity methods call phase->type(this). 115 // Make sure that the type array is big enough for 116 // our new node, even though we may throw the node away. 117 // (Note: This tweaking with igvn only works because x is a new node.) 118 _igvn.set_type(x, t); 119 // If x is a TypeNode, capture any more-precise type permanently into Node 120 // otherwise it will be not updated during igvn->transform since 121 // igvn->type(x) is set to x->Value() already. 122 x->raise_bottom_type(t); 123 Node *y = x->Identity(&_igvn); 124 if (y != x) { 125 wins++; 126 x = y; 127 } else { 128 y = _igvn.hash_find(x); 129 if (y) { 130 wins++; 131 x = y; 132 } else { 133 // Else x is a new node we are keeping 134 // We do not need register_new_node_with_optimizer 135 // because set_type has already been called. 136 _igvn._worklist.push(x); 137 } 138 } 139 } 140 if (x != the_clone && the_clone != NULL) 141 _igvn.remove_dead_node(the_clone); 142 phi->set_req( i, x ); 143 } 144 // Too few wins? 145 if (wins <= policy) { 146 _igvn.remove_dead_node(phi); 147 return NULL; 148 } 149 150 // Record Phi 151 register_new_node( phi, region ); 152 153 for (uint i2 = 1; i2 < phi->req(); i2++) { 154 Node *x = phi->in(i2); 155 // If we commoned up the cloned 'x' with another existing Node, 156 // the existing Node picks up a new use. We need to make the 157 // existing Node occur higher up so it dominates its uses. 158 Node *old_ctrl; 159 IdealLoopTree *old_loop; 160 161 if (x->is_Con()) { 162 // Constant's control is always root. 163 set_ctrl(x, C->root()); 164 continue; 165 } 166 // The occasional new node 167 if (x->_idx >= old_unique) { // Found a new, unplaced node? 168 old_ctrl = NULL; 169 old_loop = NULL; // Not in any prior loop 170 } else { 171 old_ctrl = get_ctrl(x); 172 old_loop = get_loop(old_ctrl); // Get prior loop 173 } 174 // New late point must dominate new use 175 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 176 if (new_ctrl == old_ctrl) // Nothing is changed 177 continue; 178 179 IdealLoopTree *new_loop = get_loop(new_ctrl); 180 181 // Don't move x into a loop if its uses are 182 // outside of loop. Otherwise x will be cloned 183 // for each use outside of this loop. 184 IdealLoopTree *use_loop = get_loop(region); 185 if (!new_loop->is_member(use_loop) && 186 (old_loop == NULL || !new_loop->is_member(old_loop))) { 187 // Take early control, later control will be recalculated 188 // during next iteration of loop optimizations. 189 new_ctrl = get_early_ctrl(x); 190 new_loop = get_loop(new_ctrl); 191 } 192 // Set new location 193 set_ctrl(x, new_ctrl); 194 // If changing loop bodies, see if we need to collect into new body 195 if (old_loop != new_loop) { 196 if (old_loop && !old_loop->_child) 197 old_loop->_body.yank(x); 198 if (!new_loop->_child) 199 new_loop->_body.push(x); // Collect body info 200 } 201 } 202 203 return phi; 204 } 205 206 //------------------------------dominated_by------------------------------------ 207 // Replace the dominated test with an obvious true or false. Place it on the 208 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 209 // live path up to the dominating control. 210 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 211 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 212 213 // prevdom is the dominating projection of the dominating test. 214 assert( iff->is_If(), "" ); 215 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 216 int pop = prevdom->Opcode(); 217 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 218 if (flip) { 219 if (pop == Op_IfTrue) 220 pop = Op_IfFalse; 221 else 222 pop = Op_IfTrue; 223 } 224 // 'con' is set to true or false to kill the dominated test. 225 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 226 set_ctrl(con, C->root()); // Constant gets a new use 227 // Hack the dominated test 228 _igvn.replace_input_of(iff, 1, con); 229 230 // If I dont have a reachable TRUE and FALSE path following the IfNode then 231 // I can assume this path reaches an infinite loop. In this case it's not 232 // important to optimize the data Nodes - either the whole compilation will 233 // be tossed or this path (and all data Nodes) will go dead. 234 if (iff->outcnt() != 2) return; 235 236 // Make control-dependent data Nodes on the live path (path that will remain 237 // once the dominated IF is removed) become control-dependent on the 238 // dominating projection. 239 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 240 241 // Loop predicates may have depending checks which should not 242 // be skipped. For example, range check predicate has two checks 243 // for lower and upper bounds. 244 if (dp == NULL) 245 return; 246 247 ProjNode* dp_proj = dp->as_Proj(); 248 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 249 if (exclude_loop_predicate && 250 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 251 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 252 // If this is a range check (IfNode::is_range_check), do not 253 // reorder because Compile::allow_range_check_smearing might have 254 // changed the check. 255 return; // Let IGVN transformation change control dependence. 256 } 257 258 IdealLoopTree *old_loop = get_loop(dp); 259 260 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 261 Node* cd = dp->fast_out(i); // Control-dependent node 262 if (cd->depends_only_on_test()) { 263 assert(cd->in(0) == dp, ""); 264 _igvn.replace_input_of(cd, 0, prevdom); 265 set_early_ctrl(cd); 266 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 267 if (old_loop != new_loop) { 268 if (!old_loop->_child) old_loop->_body.yank(cd); 269 if (!new_loop->_child) new_loop->_body.push(cd); 270 } 271 --i; 272 --imax; 273 } 274 } 275 } 276 277 //------------------------------has_local_phi_input---------------------------- 278 // Return TRUE if 'n' has Phi inputs from its local block and no other 279 // block-local inputs (all non-local-phi inputs come from earlier blocks) 280 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 281 Node *n_ctrl = get_ctrl(n); 282 // See if some inputs come from a Phi in this block, or from before 283 // this block. 284 uint i; 285 for( i = 1; i < n->req(); i++ ) { 286 Node *phi = n->in(i); 287 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 288 break; 289 } 290 if( i >= n->req() ) 291 return NULL; // No Phi inputs; nowhere to clone thru 292 293 // Check for inputs created between 'n' and the Phi input. These 294 // must split as well; they have already been given the chance 295 // (courtesy of a post-order visit) and since they did not we must 296 // recover the 'cost' of splitting them by being very profitable 297 // when splitting 'n'. Since this is unlikely we simply give up. 298 for( i = 1; i < n->req(); i++ ) { 299 Node *m = n->in(i); 300 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 301 // We allow the special case of AddP's with no local inputs. 302 // This allows us to split-up address expressions. 303 if (m->is_AddP() && 304 get_ctrl(m->in(2)) != n_ctrl && 305 get_ctrl(m->in(3)) != n_ctrl) { 306 // Move the AddP up to dominating point 307 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 308 continue; 309 } 310 return NULL; 311 } 312 } 313 314 return n_ctrl; 315 } 316 317 //------------------------------remix_address_expressions---------------------- 318 // Rework addressing expressions to get the most loop-invariant stuff 319 // moved out. We'd like to do all associative operators, but it's especially 320 // important (common) to do address expressions. 321 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 322 if (!has_ctrl(n)) return NULL; 323 Node *n_ctrl = get_ctrl(n); 324 IdealLoopTree *n_loop = get_loop(n_ctrl); 325 326 // See if 'n' mixes loop-varying and loop-invariant inputs and 327 // itself is loop-varying. 328 329 // Only interested in binary ops (and AddP) 330 if( n->req() < 3 || n->req() > 4 ) return NULL; 331 332 Node *n1_ctrl = get_ctrl(n->in( 1)); 333 Node *n2_ctrl = get_ctrl(n->in( 2)); 334 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 335 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 336 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 337 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 338 339 // Does one of my inputs spin in a tighter loop than self? 340 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 341 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 342 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 343 return NULL; // Leave well enough alone 344 345 // Is at least one of my inputs loop-invariant? 346 if( n1_loop == n_loop && 347 n2_loop == n_loop && 348 n3_loop == n_loop ) 349 return NULL; // No loop-invariant inputs 350 351 352 int n_op = n->Opcode(); 353 354 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 355 if( n_op == Op_LShiftI ) { 356 // Scale is loop invariant 357 Node *scale = n->in(2); 358 Node *scale_ctrl = get_ctrl(scale); 359 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 360 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 361 return NULL; 362 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 363 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 364 return NULL; // Dont bother with byte/short masking 365 // Add must vary with loop (else shift would be loop-invariant) 366 Node *add = n->in(1); 367 Node *add_ctrl = get_ctrl(add); 368 IdealLoopTree *add_loop = get_loop(add_ctrl); 369 //assert( n_loop == add_loop, "" ); 370 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 371 372 // Convert I-V into I+ (0-V); same for V-I 373 if( add->Opcode() == Op_SubI && 374 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 375 Node *zero = _igvn.intcon(0); 376 set_ctrl(zero, C->root()); 377 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 378 register_new_node( neg, get_ctrl(add->in(2) ) ); 379 add = new AddINode( add->in(1), neg ); 380 register_new_node( add, add_ctrl ); 381 } 382 if( add->Opcode() != Op_AddI ) return NULL; 383 // See if one add input is loop invariant 384 Node *add_var = add->in(1); 385 Node *add_var_ctrl = get_ctrl(add_var); 386 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 387 Node *add_invar = add->in(2); 388 Node *add_invar_ctrl = get_ctrl(add_invar); 389 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 390 if( add_var_loop == n_loop ) { 391 } else if( add_invar_loop == n_loop ) { 392 // Swap to find the invariant part 393 add_invar = add_var; 394 add_invar_ctrl = add_var_ctrl; 395 add_invar_loop = add_var_loop; 396 add_var = add->in(2); 397 Node *add_var_ctrl = get_ctrl(add_var); 398 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 399 } else // Else neither input is loop invariant 400 return NULL; 401 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 402 return NULL; // No invariant part of the add? 403 404 // Yes! Reshape address expression! 405 Node *inv_scale = new LShiftINode( add_invar, scale ); 406 Node *inv_scale_ctrl = 407 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 408 add_invar_ctrl : scale_ctrl; 409 register_new_node( inv_scale, inv_scale_ctrl ); 410 Node *var_scale = new LShiftINode( add_var, scale ); 411 register_new_node( var_scale, n_ctrl ); 412 Node *var_add = new AddINode( var_scale, inv_scale ); 413 register_new_node( var_add, n_ctrl ); 414 _igvn.replace_node( n, var_add ); 415 return var_add; 416 } 417 418 // Replace (I+V) with (V+I) 419 if( n_op == Op_AddI || 420 n_op == Op_AddL || 421 n_op == Op_AddF || 422 n_op == Op_AddD || 423 n_op == Op_MulI || 424 n_op == Op_MulL || 425 n_op == Op_MulF || 426 n_op == Op_MulD ) { 427 if( n2_loop == n_loop ) { 428 assert( n1_loop != n_loop, "" ); 429 n->swap_edges(1, 2); 430 } 431 } 432 433 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 434 // but not if I2 is a constant. 435 if( n_op == Op_AddP ) { 436 if( n2_loop == n_loop && n3_loop != n_loop ) { 437 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 438 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 439 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 440 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 441 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 442 if( n22loop != n_loop && n22loop->is_member(n_loop) && 443 n23_loop == n_loop ) { 444 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 445 // Stuff new AddP in the loop preheader 446 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 447 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 448 register_new_node( add2, n_ctrl ); 449 _igvn.replace_node( n, add2 ); 450 return add2; 451 } 452 } 453 } 454 455 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 456 if (n2_loop != n_loop && n3_loop == n_loop) { 457 if (n->in(3)->Opcode() == Op_AddX) { 458 Node *V = n->in(3)->in(1); 459 Node *I = n->in(3)->in(2); 460 if (is_member(n_loop,get_ctrl(V))) { 461 } else { 462 Node *tmp = V; V = I; I = tmp; 463 } 464 if (!is_member(n_loop,get_ctrl(I))) { 465 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 466 // Stuff new AddP in the loop preheader 467 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 468 Node *add2 = new AddPNode(n->in(1), add1, V); 469 register_new_node(add2, n_ctrl); 470 _igvn.replace_node(n, add2); 471 return add2; 472 } 473 } 474 } 475 } 476 477 return NULL; 478 } 479 480 //------------------------------conditional_move------------------------------- 481 // Attempt to replace a Phi with a conditional move. We have some pretty 482 // strict profitability requirements. All Phis at the merge point must 483 // be converted, so we can remove the control flow. We need to limit the 484 // number of c-moves to a small handful. All code that was in the side-arms 485 // of the CFG diamond is now speculatively executed. This code has to be 486 // "cheap enough". We are pretty much limited to CFG diamonds that merge 487 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 488 Node *PhaseIdealLoop::conditional_move( Node *region ) { 489 490 assert(region->is_Region(), "sanity check"); 491 if (region->req() != 3) return NULL; 492 493 // Check for CFG diamond 494 Node *lp = region->in(1); 495 Node *rp = region->in(2); 496 if (!lp || !rp) return NULL; 497 Node *lp_c = lp->in(0); 498 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 499 IfNode *iff = lp_c->as_If(); 500 501 // Check for ops pinned in an arm of the diamond. 502 // Can't remove the control flow in this case 503 if (lp->outcnt() > 1) return NULL; 504 if (rp->outcnt() > 1) return NULL; 505 506 IdealLoopTree* r_loop = get_loop(region); 507 assert(r_loop == get_loop(iff), "sanity"); 508 // Always convert to CMOVE if all results are used only outside this loop. 509 bool used_inside_loop = (r_loop == _ltree_root); 510 511 // Check profitability 512 int cost = 0; 513 int phis = 0; 514 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 515 Node *out = region->fast_out(i); 516 if (!out->is_Phi()) continue; // Ignore other control edges, etc 517 phis++; 518 PhiNode* phi = out->as_Phi(); 519 BasicType bt = phi->type()->basic_type(); 520 switch (bt) { 521 case T_DOUBLE: 522 if (C->use_cmove()) { 523 continue; //TODO: maybe we want to add some cost 524 } 525 case T_FLOAT: { 526 cost += Matcher::float_cmove_cost(); // Could be very expensive 527 break; 528 } 529 case T_LONG: { 530 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 531 } 532 case T_INT: // These all CMOV fine 533 case T_ADDRESS: { // (RawPtr) 534 cost++; 535 break; 536 } 537 case T_NARROWOOP: // Fall through 538 case T_OBJECT: { // Base oops are OK, but not derived oops 539 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 540 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 541 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 542 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 543 // have a Phi for the base here that we convert to a CMOVE all is well 544 // and good. But if the base is dead, we'll not make a CMOVE. Later 545 // the allocator will have to produce a base by creating a CMOVE of the 546 // relevant bases. This puts the allocator in the business of 547 // manufacturing expensive instructions, generally a bad plan. 548 // Just Say No to Conditionally-Moved Derived Pointers. 549 if (tp && tp->offset() != 0) 550 return NULL; 551 cost++; 552 break; 553 } 554 default: 555 return NULL; // In particular, can't do memory or I/O 556 } 557 // Add in cost any speculative ops 558 for (uint j = 1; j < region->req(); j++) { 559 Node *proj = region->in(j); 560 Node *inp = phi->in(j); 561 if (get_ctrl(inp) == proj) { // Found local op 562 cost++; 563 // Check for a chain of dependent ops; these will all become 564 // speculative in a CMOV. 565 for (uint k = 1; k < inp->req(); k++) 566 if (get_ctrl(inp->in(k)) == proj) 567 cost += ConditionalMoveLimit; // Too much speculative goo 568 } 569 } 570 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 571 // This will likely Split-If, a higher-payoff operation. 572 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 573 Node* use = phi->fast_out(k); 574 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 575 cost += ConditionalMoveLimit; 576 // Is there a use inside the loop? 577 // Note: check only basic types since CMoveP is pinned. 578 if (!used_inside_loop && is_java_primitive(bt)) { 579 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 580 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 581 used_inside_loop = true; 582 } 583 } 584 } 585 }//for 586 Node* bol = iff->in(1); 587 assert(bol->Opcode() == Op_Bool, ""); 588 int cmp_op = bol->in(1)->Opcode(); 589 // It is expensive to generate flags from a float compare. 590 // Avoid duplicated float compare. 591 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 592 593 float infrequent_prob = PROB_UNLIKELY_MAG(3); 594 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 595 if (used_inside_loop) { 596 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 597 598 // BlockLayoutByFrequency optimization moves infrequent branch 599 // from hot path. No point in CMOV'ing in such case (110 is used 600 // instead of 100 to take into account not exactness of float value). 601 if (BlockLayoutByFrequency) { 602 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 603 } 604 } 605 // Check for highly predictable branch. No point in CMOV'ing if 606 // we are going to predict accurately all the time. 607 if (C->use_cmove() && cmp_op == Op_CmpD) ;//keep going 608 else if (iff->_prob < infrequent_prob || 609 iff->_prob > (1.0f - infrequent_prob)) 610 return NULL; 611 612 // -------------- 613 // Now replace all Phis with CMOV's 614 Node *cmov_ctrl = iff->in(0); 615 uint flip = (lp->Opcode() == Op_IfTrue); 616 while (1) { 617 PhiNode* phi = NULL; 618 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 619 Node *out = region->fast_out(i); 620 if (out->is_Phi()) { 621 phi = out->as_Phi(); 622 break; 623 } 624 } 625 if (phi == NULL) break; 626 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 627 // Move speculative ops 628 for (uint j = 1; j < region->req(); j++) { 629 Node *proj = region->in(j); 630 Node *inp = phi->in(j); 631 if (get_ctrl(inp) == proj) { // Found local op 632 #ifndef PRODUCT 633 if (PrintOpto && VerifyLoopOptimizations) { 634 tty->print(" speculate: "); 635 inp->dump(); 636 } 637 #endif 638 set_ctrl(inp, cmov_ctrl); 639 } 640 } 641 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 642 register_new_node( cmov, cmov_ctrl ); 643 _igvn.replace_node( phi, cmov ); 644 #ifndef PRODUCT 645 if (TraceLoopOpts) { 646 tty->print("CMOV "); 647 r_loop->dump_head(); 648 if (Verbose) { 649 bol->in(1)->dump(1); 650 cmov->dump(1); 651 } 652 } 653 if (VerifyLoopOptimizations) verify(); 654 #endif 655 } 656 657 // The useless CFG diamond will fold up later; see the optimization in 658 // RegionNode::Ideal. 659 _igvn._worklist.push(region); 660 661 return iff->in(1); 662 } 663 664 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 665 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 666 Node* u = m->fast_out(i); 667 if (u->is_CFG()) { 668 if (u->Opcode() == Op_NeverBranch) { 669 u = ((NeverBranchNode*)u)->proj_out(0); 670 enqueue_cfg_uses(u, wq); 671 } else { 672 wq.push(u); 673 } 674 } 675 } 676 } 677 678 // Try moving a store out of a loop, right before the loop 679 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 680 // Store has to be first in the loop body 681 IdealLoopTree *n_loop = get_loop(n_ctrl); 682 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) { 683 Node* address = n->in(MemNode::Address); 684 Node* value = n->in(MemNode::ValueIn); 685 Node* mem = n->in(MemNode::Memory); 686 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 687 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 688 689 // - address and value must be loop invariant 690 // - memory must be a memory Phi for the loop 691 // - Store must be the only store on this memory slice in the 692 // loop: if there's another store following this one then value 693 // written at iteration i by the second store could be overwritten 694 // at iteration i+n by the first store: it's not safe to move the 695 // first store out of the loop 696 // - nothing must observe the memory Phi: it guarantees no read 697 // before the store, we are also guaranteed the store post 698 // dominates the loop head (ignoring a possible early 699 // exit). Otherwise there would be extra Phi involved between the 700 // loop's Phi and the store. 701 // - there must be no early exit from the loop before the Store 702 // (such an exit most of the time would be an extra use of the 703 // memory Phi but sometimes is a bottom memory Phi that takes the 704 // store as input). 705 706 if (!n_loop->is_member(address_loop) && 707 !n_loop->is_member(value_loop) && 708 mem->is_Phi() && mem->in(0) == n_loop->_head && 709 mem->outcnt() == 1 && 710 mem->in(LoopNode::LoopBackControl) == n) { 711 712 assert(n_loop->_tail != NULL, "need a tail"); 713 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 714 715 // Verify that there's no early exit of the loop before the store. 716 bool ctrl_ok = false; 717 { 718 // Follow control from loop head until n, we exit the loop or 719 // we reach the tail 720 ResourceMark rm; 721 Unique_Node_List wq; 722 wq.push(n_loop->_head); 723 724 for (uint next = 0; next < wq.size(); ++next) { 725 Node *m = wq.at(next); 726 if (m == n->in(0)) { 727 ctrl_ok = true; 728 continue; 729 } 730 assert(!has_ctrl(m), "should be CFG"); 731 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 732 ctrl_ok = false; 733 break; 734 } 735 enqueue_cfg_uses(m, wq); 736 if (wq.size() > 10) { 737 ctrl_ok = false; 738 break; 739 } 740 } 741 } 742 if (ctrl_ok) { 743 // move the Store 744 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 745 _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl)); 746 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 747 // Disconnect the phi now. An empty phi can confuse other 748 // optimizations in this pass of loop opts. 749 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 750 n_loop->_body.yank(mem); 751 752 IdealLoopTree* new_loop = get_loop(n->in(0)); 753 set_ctrl_and_loop(n, n->in(0)); 754 755 return n; 756 } 757 } 758 } 759 return NULL; 760 } 761 762 // Try moving a store out of a loop, right after the loop 763 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 764 if (n->is_Store() && n->in(0) != NULL) { 765 Node *n_ctrl = get_ctrl(n); 766 IdealLoopTree *n_loop = get_loop(n_ctrl); 767 // Store must be in a loop 768 if (n_loop != _ltree_root && !n_loop->_irreducible) { 769 Node* address = n->in(MemNode::Address); 770 Node* value = n->in(MemNode::ValueIn); 771 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 772 // address must be loop invariant 773 if (!n_loop->is_member(address_loop)) { 774 // Store must be last on this memory slice in the loop and 775 // nothing in the loop must observe it 776 Node* phi = NULL; 777 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 778 Node* u = n->fast_out(i); 779 if (has_ctrl(u)) { // control use? 780 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 781 if (!n_loop->is_member(u_loop)) { 782 continue; 783 } 784 if (u->is_Phi() && u->in(0) == n_loop->_head) { 785 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 786 // multiple phis on the same slice are possible 787 if (phi != NULL) { 788 return; 789 } 790 phi = u; 791 continue; 792 } 793 } 794 return; 795 } 796 if (phi != NULL) { 797 // Nothing in the loop before the store (next iteration) 798 // must observe the stored value 799 bool mem_ok = true; 800 { 801 ResourceMark rm; 802 Unique_Node_List wq; 803 wq.push(phi); 804 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 805 Node *m = wq.at(next); 806 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 807 Node* u = m->fast_out(i); 808 if (u->is_Store() || u->is_Phi()) { 809 if (u != n) { 810 wq.push(u); 811 mem_ok = (wq.size() <= 10); 812 } 813 } else { 814 mem_ok = false; 815 break; 816 } 817 } 818 } 819 } 820 if (mem_ok) { 821 // Move the Store out of the loop creating clones along 822 // all paths out of the loop that observe the stored value 823 _igvn.rehash_node_delayed(phi); 824 int count = phi->replace_edge(n, n->in(MemNode::Memory)); 825 assert(count > 0, "inconsistent phi"); 826 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 827 Node* u = n->fast_out(i); 828 Node* c = get_ctrl(u); 829 830 if (u->is_Phi()) { 831 c = u->in(0)->in(u->find_edge(n)); 832 } 833 IdealLoopTree *u_loop = get_loop(c); 834 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop"); 835 while(true) { 836 Node* next_c = find_non_split_ctrl(idom(c)); 837 if (n_loop->is_member(get_loop(next_c))) { 838 break; 839 } 840 c = next_c; 841 } 842 843 Node* st = n->clone(); 844 st->set_req(0, c); 845 _igvn.register_new_node_with_optimizer(st); 846 847 set_ctrl(st, c); 848 IdealLoopTree* new_loop = get_loop(c); 849 assert(new_loop != n_loop, "should be moved out of loop"); 850 if (new_loop->_child == NULL) new_loop->_body.push(st); 851 852 _igvn.replace_input_of(u, u->find_edge(n), st); 853 --imax; 854 --i; 855 } 856 857 858 assert(n->outcnt() == 0, "all uses should be gone"); 859 _igvn.replace_input_of(n, MemNode::Memory, C->top()); 860 // Disconnect the phi now. An empty phi can confuse other 861 // optimizations in this pass of loop opts.. 862 if (phi->in(LoopNode::LoopBackControl) == phi) { 863 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 864 n_loop->_body.yank(phi); 865 } 866 } 867 } 868 } 869 } 870 } 871 } 872 873 //------------------------------split_if_with_blocks_pre----------------------- 874 // Do the real work in a non-recursive function. Data nodes want to be 875 // cloned in the pre-order so they can feed each other nicely. 876 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 877 // Cloning these guys is unlikely to win 878 int n_op = n->Opcode(); 879 if( n_op == Op_MergeMem ) return n; 880 if( n->is_Proj() ) return n; 881 // Do not clone-up CmpFXXX variations, as these are always 882 // followed by a CmpI 883 if( n->is_Cmp() ) return n; 884 // Attempt to use a conditional move instead of a phi/branch 885 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 886 Node *cmov = conditional_move( n ); 887 if( cmov ) return cmov; 888 } 889 if( n->is_CFG() || n->is_LoadStore() ) 890 return n; 891 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 892 n_op == Op_Opaque2 ) { 893 if( !C->major_progress() ) // If chance of no more loop opts... 894 _igvn._worklist.push(n); // maybe we'll remove them 895 return n; 896 } 897 898 if( n->is_Con() ) return n; // No cloning for Con nodes 899 900 Node *n_ctrl = get_ctrl(n); 901 if( !n_ctrl ) return n; // Dead node 902 903 Node* res = try_move_store_before_loop(n, n_ctrl); 904 if (res != NULL) { 905 return n; 906 } 907 908 // Attempt to remix address expressions for loop invariants 909 Node *m = remix_address_expressions( n ); 910 if( m ) return m; 911 912 if (n->is_ConstraintCast()) { 913 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(this); 914 if (dom_cast != NULL) { 915 _igvn.replace_node(n, dom_cast); 916 return dom_cast; 917 } 918 } 919 920 // Determine if the Node has inputs from some local Phi. 921 // Returns the block to clone thru. 922 Node *n_blk = has_local_phi_input( n ); 923 if( !n_blk ) return n; 924 925 // Do not clone the trip counter through on a CountedLoop 926 // (messes up the canonical shape). 927 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 928 929 // Check for having no control input; not pinned. Allow 930 // dominating control. 931 if (n->in(0)) { 932 Node *dom = idom(n_blk); 933 if (dom_lca(n->in(0), dom) != n->in(0)) { 934 return n; 935 } 936 } 937 // Policy: when is it profitable. You must get more wins than 938 // policy before it is considered profitable. Policy is usually 0, 939 // so 1 win is considered profitable. Big merges will require big 940 // cloning, so get a larger policy. 941 int policy = n_blk->req() >> 2; 942 943 // If the loop is a candidate for range check elimination, 944 // delay splitting through it's phi until a later loop optimization 945 if (n_blk->is_CountedLoop()) { 946 IdealLoopTree *lp = get_loop(n_blk); 947 if (lp && lp->_rce_candidate) { 948 return n; 949 } 950 } 951 952 // Use same limit as split_if_with_blocks_post 953 if( C->live_nodes() > 35000 ) return n; // Method too big 954 955 // Split 'n' through the merge point if it is profitable 956 Node *phi = split_thru_phi( n, n_blk, policy ); 957 if (!phi) return n; 958 959 // Found a Phi to split thru! 960 // Replace 'n' with the new phi 961 _igvn.replace_node( n, phi ); 962 // Moved a load around the loop, 'en-registering' something. 963 if (n_blk->is_Loop() && n->is_Load() && 964 !phi->in(LoopNode::LoopBackControl)->is_Load()) 965 C->set_major_progress(); 966 967 return phi; 968 } 969 970 static bool merge_point_too_heavy(Compile* C, Node* region) { 971 // Bail out if the region and its phis have too many users. 972 int weight = 0; 973 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 974 weight += region->fast_out(i)->outcnt(); 975 } 976 int nodes_left = C->max_node_limit() - C->live_nodes(); 977 if (weight * 8 > nodes_left) { 978 if (PrintOpto) { 979 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 980 } 981 return true; 982 } else { 983 return false; 984 } 985 } 986 987 static bool merge_point_safe(Node* region) { 988 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 989 // having a PhiNode input. This sidesteps the dangerous case where the split 990 // ConvI2LNode may become TOP if the input Value() does not 991 // overlap the ConvI2L range, leaving a node which may not dominate its 992 // uses. 993 // A better fix for this problem can be found in the BugTraq entry, but 994 // expediency for Mantis demands this hack. 995 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 996 // split_if_with_blocks from splitting a block because we could not move around 997 // the FastLockNode. 998 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 999 Node* n = region->fast_out(i); 1000 if (n->is_Phi()) { 1001 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1002 Node* m = n->fast_out(j); 1003 if (m->is_FastLock()) 1004 return false; 1005 #ifdef _LP64 1006 if (m->Opcode() == Op_ConvI2L) 1007 return false; 1008 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1009 return false; 1010 } 1011 #endif 1012 } 1013 } 1014 } 1015 return true; 1016 } 1017 1018 1019 //------------------------------place_near_use--------------------------------- 1020 // Place some computation next to use but not inside inner loops. 1021 // For inner loop uses move it to the preheader area. 1022 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1023 IdealLoopTree *u_loop = get_loop( useblock ); 1024 return (u_loop->_irreducible || u_loop->_child) 1025 ? useblock 1026 : u_loop->_head->in(LoopNode::EntryControl); 1027 } 1028 1029 1030 //------------------------------split_if_with_blocks_post---------------------- 1031 // Do the real work in a non-recursive function. CFG hackery wants to be 1032 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1033 // info. 1034 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 1035 1036 // Cloning Cmp through Phi's involves the split-if transform. 1037 // FastLock is not used by an If 1038 if( n->is_Cmp() && !n->is_FastLock() ) { 1039 if( C->live_nodes() > 35000 ) return; // Method too big 1040 1041 // Do not do 'split-if' if irreducible loops are present. 1042 if( _has_irreducible_loops ) 1043 return; 1044 1045 Node *n_ctrl = get_ctrl(n); 1046 // Determine if the Node has inputs from some local Phi. 1047 // Returns the block to clone thru. 1048 Node *n_blk = has_local_phi_input( n ); 1049 if( n_blk != n_ctrl ) return; 1050 1051 if( merge_point_too_heavy(C, n_ctrl) ) 1052 return; 1053 1054 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 1055 Node *bol = n->unique_out(); 1056 assert( bol->is_Bool(), "expect a bool here" ); 1057 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 1058 Node *iff = bol->unique_out(); 1059 1060 // Check some safety conditions 1061 if( iff->is_If() ) { // Classic split-if? 1062 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 1063 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1064 // Can't split CMove with different control edge. 1065 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return; 1066 if( get_ctrl(iff->in(2)) == n_ctrl || 1067 get_ctrl(iff->in(3)) == n_ctrl ) 1068 return; // Inputs not yet split-up 1069 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 1070 return; // Loop-invar test gates loop-varying CMOVE 1071 } 1072 } else { 1073 return; // some other kind of node, such as an Allocate 1074 } 1075 1076 // Do not do 'split-if' if some paths are dead. First do dead code 1077 // elimination and then see if its still profitable. 1078 for( uint i = 1; i < n_ctrl->req(); i++ ) 1079 if( n_ctrl->in(i) == C->top() ) 1080 return; 1081 1082 // When is split-if profitable? Every 'win' on means some control flow 1083 // goes dead, so it's almost always a win. 1084 int policy = 0; 1085 // If trying to do a 'Split-If' at the loop head, it is only 1086 // profitable if the cmp folds up on BOTH paths. Otherwise we 1087 // risk peeling a loop forever. 1088 1089 // CNC - Disabled for now. Requires careful handling of loop 1090 // body selection for the cloned code. Also, make sure we check 1091 // for any input path not being in the same loop as n_ctrl. For 1092 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1093 // because the alternative loop entry points won't be converted 1094 // into LoopNodes. 1095 IdealLoopTree *n_loop = get_loop(n_ctrl); 1096 for( uint j = 1; j < n_ctrl->req(); j++ ) 1097 if( get_loop(n_ctrl->in(j)) != n_loop ) 1098 return; 1099 1100 // Check for safety of the merge point. 1101 if( !merge_point_safe(n_ctrl) ) { 1102 return; 1103 } 1104 1105 // Split compare 'n' through the merge point if it is profitable 1106 Node *phi = split_thru_phi( n, n_ctrl, policy ); 1107 if( !phi ) return; 1108 1109 // Found a Phi to split thru! 1110 // Replace 'n' with the new phi 1111 _igvn.replace_node( n, phi ); 1112 1113 // Now split the bool up thru the phi 1114 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 1115 guarantee(bolphi != NULL, "null boolean phi node"); 1116 1117 _igvn.replace_node( bol, bolphi ); 1118 assert( iff->in(1) == bolphi, "" ); 1119 1120 if( bolphi->Value(&_igvn)->singleton() ) 1121 return; 1122 1123 // Conditional-move? Must split up now 1124 if( !iff->is_If() ) { 1125 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 1126 _igvn.replace_node( iff, cmovphi ); 1127 return; 1128 } 1129 1130 // Now split the IF 1131 do_split_if( iff ); 1132 return; 1133 } 1134 1135 // Check for an IF ready to split; one that has its 1136 // condition codes input coming from a Phi at the block start. 1137 int n_op = n->Opcode(); 1138 1139 // Check for an IF being dominated by another IF same test 1140 if (n_op == Op_If || 1141 n_op == Op_RangeCheck) { 1142 Node *bol = n->in(1); 1143 uint max = bol->outcnt(); 1144 // Check for same test used more than once? 1145 if (max > 1 && bol->is_Bool()) { 1146 // Search up IDOMs to see if this IF is dominated. 1147 Node *cutoff = get_ctrl(bol); 1148 1149 // Now search up IDOMs till cutoff, looking for a dominating test 1150 Node *prevdom = n; 1151 Node *dom = idom(prevdom); 1152 while (dom != cutoff) { 1153 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1154 // Replace the dominated test with an obvious true or false. 1155 // Place it on the IGVN worklist for later cleanup. 1156 C->set_major_progress(); 1157 dominated_by(prevdom, n, false, true); 1158 #ifndef PRODUCT 1159 if( VerifyLoopOptimizations ) verify(); 1160 #endif 1161 return; 1162 } 1163 prevdom = dom; 1164 dom = idom(prevdom); 1165 } 1166 } 1167 } 1168 1169 // See if a shared loop-varying computation has no loop-varying uses. 1170 // Happens if something is only used for JVM state in uncommon trap exits, 1171 // like various versions of induction variable+offset. Clone the 1172 // computation per usage to allow it to sink out of the loop. 1173 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1174 Node *n_ctrl = get_ctrl(n); 1175 IdealLoopTree *n_loop = get_loop(n_ctrl); 1176 if( n_loop != _ltree_root ) { 1177 DUIterator_Fast imax, i = n->fast_outs(imax); 1178 for (; i < imax; i++) { 1179 Node* u = n->fast_out(i); 1180 if( !has_ctrl(u) ) break; // Found control user 1181 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1182 if( u_loop == n_loop ) break; // Found loop-varying use 1183 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1184 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1185 } 1186 bool did_break = (i < imax); // Did we break out of the previous loop? 1187 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1188 Node *late_load_ctrl = NULL; 1189 if (n->is_Load()) { 1190 // If n is a load, get and save the result from get_late_ctrl(), 1191 // to be later used in calculating the control for n's clones. 1192 clear_dom_lca_tags(); 1193 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1194 } 1195 // If n is a load, and the late control is the same as the current 1196 // control, then the cloning of n is a pointless exercise, because 1197 // GVN will ensure that we end up where we started. 1198 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1199 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1200 Node *u = n->last_out(j); // Clone private computation per use 1201 _igvn.rehash_node_delayed(u); 1202 Node *x = n->clone(); // Clone computation 1203 Node *x_ctrl = NULL; 1204 if( u->is_Phi() ) { 1205 // Replace all uses of normal nodes. Replace Phi uses 1206 // individually, so the separate Nodes can sink down 1207 // different paths. 1208 uint k = 1; 1209 while( u->in(k) != n ) k++; 1210 u->set_req( k, x ); 1211 // x goes next to Phi input path 1212 x_ctrl = u->in(0)->in(k); 1213 --j; 1214 } else { // Normal use 1215 // Replace all uses 1216 for( uint k = 0; k < u->req(); k++ ) { 1217 if( u->in(k) == n ) { 1218 u->set_req( k, x ); 1219 --j; 1220 } 1221 } 1222 x_ctrl = get_ctrl(u); 1223 } 1224 1225 // Find control for 'x' next to use but not inside inner loops. 1226 // For inner loop uses get the preheader area. 1227 x_ctrl = place_near_use(x_ctrl); 1228 1229 if (n->is_Load()) { 1230 // For loads, add a control edge to a CFG node outside of the loop 1231 // to force them to not combine and return back inside the loop 1232 // during GVN optimization (4641526). 1233 // 1234 // Because we are setting the actual control input, factor in 1235 // the result from get_late_ctrl() so we respect any 1236 // anti-dependences. (6233005). 1237 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1238 1239 // Don't allow the control input to be a CFG splitting node. 1240 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1241 // should only have IfTrueNode and IfFalseNode (4985384). 1242 x_ctrl = find_non_split_ctrl(x_ctrl); 1243 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1244 1245 x->set_req(0, x_ctrl); 1246 } 1247 register_new_node(x, x_ctrl); 1248 1249 // Some institutional knowledge is needed here: 'x' is 1250 // yanked because if the optimizer runs GVN on it all the 1251 // cloned x's will common up and undo this optimization and 1252 // be forced back in the loop. This is annoying because it 1253 // makes +VerifyOpto report false-positives on progress. I 1254 // tried setting control edges on the x's to force them to 1255 // not combine, but the matching gets worried when it tries 1256 // to fold a StoreP and an AddP together (as part of an 1257 // address expression) and the AddP and StoreP have 1258 // different controls. 1259 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1260 } 1261 _igvn.remove_dead_node(n); 1262 } 1263 } 1264 } 1265 } 1266 1267 try_move_store_after_loop(n); 1268 1269 // Check for Opaque2's who's loop has disappeared - who's input is in the 1270 // same loop nest as their output. Remove 'em, they are no longer useful. 1271 if( n_op == Op_Opaque2 && 1272 n->in(1) != NULL && 1273 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1274 _igvn.replace_node( n, n->in(1) ); 1275 } 1276 } 1277 1278 //------------------------------split_if_with_blocks--------------------------- 1279 // Check for aggressive application of 'split-if' optimization, 1280 // using basic block level info. 1281 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1282 Node *n = C->root(); 1283 visited.set(n->_idx); // first, mark node as visited 1284 // Do pre-visit work for root 1285 n = split_if_with_blocks_pre( n ); 1286 uint cnt = n->outcnt(); 1287 uint i = 0; 1288 while (true) { 1289 // Visit all children 1290 if (i < cnt) { 1291 Node* use = n->raw_out(i); 1292 ++i; 1293 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1294 // Now do pre-visit work for this use 1295 use = split_if_with_blocks_pre( use ); 1296 nstack.push(n, i); // Save parent and next use's index. 1297 n = use; // Process all children of current use. 1298 cnt = use->outcnt(); 1299 i = 0; 1300 } 1301 } 1302 else { 1303 // All of n's children have been processed, complete post-processing. 1304 if (cnt != 0 && !n->is_Con()) { 1305 assert(has_node(n), "no dead nodes"); 1306 split_if_with_blocks_post( n ); 1307 } 1308 if (nstack.is_empty()) { 1309 // Finished all nodes on stack. 1310 break; 1311 } 1312 // Get saved parent node and next use's index. Visit the rest of uses. 1313 n = nstack.node(); 1314 cnt = n->outcnt(); 1315 i = nstack.index(); 1316 nstack.pop(); 1317 } 1318 } 1319 } 1320 1321 1322 //============================================================================= 1323 // 1324 // C L O N E A L O O P B O D Y 1325 // 1326 1327 //------------------------------clone_iff-------------------------------------- 1328 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1329 // "Nearly" because all Nodes have been cloned from the original in the loop, 1330 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1331 // through the Phi recursively, and return a Bool. 1332 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1333 1334 // Convert this Phi into a Phi merging Bools 1335 uint i; 1336 for( i = 1; i < phi->req(); i++ ) { 1337 Node *b = phi->in(i); 1338 if( b->is_Phi() ) { 1339 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1340 } else { 1341 assert( b->is_Bool(), "" ); 1342 } 1343 } 1344 1345 Node *sample_bool = phi->in(1); 1346 Node *sample_cmp = sample_bool->in(1); 1347 1348 // Make Phis to merge the Cmp's inputs. 1349 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1350 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1351 for( i = 1; i < phi->req(); i++ ) { 1352 Node *n1 = phi->in(i)->in(1)->in(1); 1353 Node *n2 = phi->in(i)->in(1)->in(2); 1354 phi1->set_req( i, n1 ); 1355 phi2->set_req( i, n2 ); 1356 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1357 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1358 } 1359 // See if these Phis have been made before. 1360 // Register with optimizer 1361 Node *hit1 = _igvn.hash_find_insert(phi1); 1362 if( hit1 ) { // Hit, toss just made Phi 1363 _igvn.remove_dead_node(phi1); // Remove new phi 1364 assert( hit1->is_Phi(), "" ); 1365 phi1 = (PhiNode*)hit1; // Use existing phi 1366 } else { // Miss 1367 _igvn.register_new_node_with_optimizer(phi1); 1368 } 1369 Node *hit2 = _igvn.hash_find_insert(phi2); 1370 if( hit2 ) { // Hit, toss just made Phi 1371 _igvn.remove_dead_node(phi2); // Remove new phi 1372 assert( hit2->is_Phi(), "" ); 1373 phi2 = (PhiNode*)hit2; // Use existing phi 1374 } else { // Miss 1375 _igvn.register_new_node_with_optimizer(phi2); 1376 } 1377 // Register Phis with loop/block info 1378 set_ctrl(phi1, phi->in(0)); 1379 set_ctrl(phi2, phi->in(0)); 1380 // Make a new Cmp 1381 Node *cmp = sample_cmp->clone(); 1382 cmp->set_req( 1, phi1 ); 1383 cmp->set_req( 2, phi2 ); 1384 _igvn.register_new_node_with_optimizer(cmp); 1385 set_ctrl(cmp, phi->in(0)); 1386 1387 // Make a new Bool 1388 Node *b = sample_bool->clone(); 1389 b->set_req(1,cmp); 1390 _igvn.register_new_node_with_optimizer(b); 1391 set_ctrl(b, phi->in(0)); 1392 1393 assert( b->is_Bool(), "" ); 1394 return (BoolNode*)b; 1395 } 1396 1397 //------------------------------clone_bool------------------------------------- 1398 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1399 // "Nearly" because all Nodes have been cloned from the original in the loop, 1400 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1401 // through the Phi recursively, and return a Bool. 1402 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1403 uint i; 1404 // Convert this Phi into a Phi merging Bools 1405 for( i = 1; i < phi->req(); i++ ) { 1406 Node *b = phi->in(i); 1407 if( b->is_Phi() ) { 1408 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1409 } else { 1410 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1411 } 1412 } 1413 1414 Node *sample_cmp = phi->in(1); 1415 1416 // Make Phis to merge the Cmp's inputs. 1417 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1418 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1419 for( uint j = 1; j < phi->req(); j++ ) { 1420 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1421 Node *n1, *n2; 1422 if( cmp_top->is_Cmp() ) { 1423 n1 = cmp_top->in(1); 1424 n2 = cmp_top->in(2); 1425 } else { 1426 n1 = n2 = cmp_top; 1427 } 1428 phi1->set_req( j, n1 ); 1429 phi2->set_req( j, n2 ); 1430 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1431 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1432 } 1433 1434 // See if these Phis have been made before. 1435 // Register with optimizer 1436 Node *hit1 = _igvn.hash_find_insert(phi1); 1437 if( hit1 ) { // Hit, toss just made Phi 1438 _igvn.remove_dead_node(phi1); // Remove new phi 1439 assert( hit1->is_Phi(), "" ); 1440 phi1 = (PhiNode*)hit1; // Use existing phi 1441 } else { // Miss 1442 _igvn.register_new_node_with_optimizer(phi1); 1443 } 1444 Node *hit2 = _igvn.hash_find_insert(phi2); 1445 if( hit2 ) { // Hit, toss just made Phi 1446 _igvn.remove_dead_node(phi2); // Remove new phi 1447 assert( hit2->is_Phi(), "" ); 1448 phi2 = (PhiNode*)hit2; // Use existing phi 1449 } else { // Miss 1450 _igvn.register_new_node_with_optimizer(phi2); 1451 } 1452 // Register Phis with loop/block info 1453 set_ctrl(phi1, phi->in(0)); 1454 set_ctrl(phi2, phi->in(0)); 1455 // Make a new Cmp 1456 Node *cmp = sample_cmp->clone(); 1457 cmp->set_req( 1, phi1 ); 1458 cmp->set_req( 2, phi2 ); 1459 _igvn.register_new_node_with_optimizer(cmp); 1460 set_ctrl(cmp, phi->in(0)); 1461 1462 assert( cmp->is_Cmp(), "" ); 1463 return (CmpNode*)cmp; 1464 } 1465 1466 //------------------------------sink_use--------------------------------------- 1467 // If 'use' was in the loop-exit block, it now needs to be sunk 1468 // below the post-loop merge point. 1469 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1470 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1471 set_ctrl(use, post_loop); 1472 for (DUIterator j = use->outs(); use->has_out(j); j++) 1473 sink_use(use->out(j), post_loop); 1474 } 1475 } 1476 1477 //------------------------------clone_loop------------------------------------- 1478 // 1479 // C L O N E A L O O P B O D Y 1480 // 1481 // This is the basic building block of the loop optimizations. It clones an 1482 // entire loop body. It makes an old_new loop body mapping; with this mapping 1483 // you can find the new-loop equivalent to an old-loop node. All new-loop 1484 // nodes are exactly equal to their old-loop counterparts, all edges are the 1485 // same. All exits from the old-loop now have a RegionNode that merges the 1486 // equivalent new-loop path. This is true even for the normal "loop-exit" 1487 // condition. All uses of loop-invariant old-loop values now come from (one 1488 // or more) Phis that merge their new-loop equivalents. 1489 // 1490 // This operation leaves the graph in an illegal state: there are two valid 1491 // control edges coming from the loop pre-header to both loop bodies. I'll 1492 // definitely have to hack the graph after running this transform. 1493 // 1494 // From this building block I will further edit edges to perform loop peeling 1495 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1496 // 1497 // Parameter side_by_size_idom: 1498 // When side_by_size_idom is NULL, the dominator tree is constructed for 1499 // the clone loop to dominate the original. Used in construction of 1500 // pre-main-post loop sequence. 1501 // When nonnull, the clone and original are side-by-side, both are 1502 // dominated by the side_by_side_idom node. Used in construction of 1503 // unswitched loops. 1504 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1505 Node* side_by_side_idom) { 1506 1507 if (C->do_vector_loop() && PrintOpto) { 1508 const char* mname = C->method()->name()->as_quoted_ascii(); 1509 if (mname != NULL) { 1510 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1511 } 1512 } 1513 1514 CloneMap& cm = C->clone_map(); 1515 Dict* dict = cm.dict(); 1516 if (C->do_vector_loop()) { 1517 cm.set_clone_idx(cm.max_gen()+1); 1518 #ifndef PRODUCT 1519 if (PrintOpto) { 1520 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1521 loop->dump_head(); 1522 } 1523 #endif 1524 } 1525 1526 // Step 1: Clone the loop body. Make the old->new mapping. 1527 uint i; 1528 for( i = 0; i < loop->_body.size(); i++ ) { 1529 Node *old = loop->_body.at(i); 1530 Node *nnn = old->clone(); 1531 old_new.map( old->_idx, nnn ); 1532 if (C->do_vector_loop()) { 1533 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1534 } 1535 _igvn.register_new_node_with_optimizer(nnn); 1536 } 1537 1538 1539 // Step 2: Fix the edges in the new body. If the old input is outside the 1540 // loop use it. If the old input is INside the loop, use the corresponding 1541 // new node instead. 1542 for( i = 0; i < loop->_body.size(); i++ ) { 1543 Node *old = loop->_body.at(i); 1544 Node *nnn = old_new[old->_idx]; 1545 // Fix CFG/Loop controlling the new node 1546 if (has_ctrl(old)) { 1547 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1548 } else { 1549 set_loop(nnn, loop->_parent); 1550 if (old->outcnt() > 0) { 1551 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1552 } 1553 } 1554 // Correct edges to the new node 1555 for( uint j = 0; j < nnn->req(); j++ ) { 1556 Node *n = nnn->in(j); 1557 if( n ) { 1558 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1559 if( loop->is_member( old_in_loop ) ) 1560 nnn->set_req(j, old_new[n->_idx]); 1561 } 1562 } 1563 _igvn.hash_find_insert(nnn); 1564 } 1565 Node *newhead = old_new[loop->_head->_idx]; 1566 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1567 1568 1569 // Step 3: Now fix control uses. Loop varying control uses have already 1570 // been fixed up (as part of all input edges in Step 2). Loop invariant 1571 // control uses must be either an IfFalse or an IfTrue. Make a merge 1572 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1573 // refer to this. 1574 ResourceArea *area = Thread::current()->resource_area(); 1575 Node_List worklist(area); 1576 uint new_counter = C->unique(); 1577 for( i = 0; i < loop->_body.size(); i++ ) { 1578 Node* old = loop->_body.at(i); 1579 if( !old->is_CFG() ) continue; 1580 Node* nnn = old_new[old->_idx]; 1581 1582 // Copy uses to a worklist, so I can munge the def-use info 1583 // with impunity. 1584 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1585 worklist.push(old->fast_out(j)); 1586 1587 while( worklist.size() ) { // Visit all uses 1588 Node *use = worklist.pop(); 1589 if (!has_node(use)) continue; // Ignore dead nodes 1590 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1591 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1592 // Both OLD and USE are CFG nodes here. 1593 assert( use->is_Proj(), "" ); 1594 1595 // Clone the loop exit control projection 1596 Node *newuse = use->clone(); 1597 if (C->do_vector_loop()) { 1598 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1599 } 1600 newuse->set_req(0,nnn); 1601 _igvn.register_new_node_with_optimizer(newuse); 1602 set_loop(newuse, use_loop); 1603 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1604 1605 // We need a Region to merge the exit from the peeled body and the 1606 // exit from the old loop body. 1607 RegionNode *r = new RegionNode(3); 1608 // Map the old use to the new merge point 1609 old_new.map( use->_idx, r ); 1610 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1611 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1612 1613 // The original user of 'use' uses 'r' instead. 1614 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1615 Node* useuse = use->last_out(l); 1616 _igvn.rehash_node_delayed(useuse); 1617 uint uses_found = 0; 1618 if( useuse->in(0) == use ) { 1619 useuse->set_req(0, r); 1620 uses_found++; 1621 if( useuse->is_CFG() ) { 1622 assert( dom_depth(useuse) > dd_r, "" ); 1623 set_idom(useuse, r, dom_depth(useuse)); 1624 } 1625 } 1626 for( uint k = 1; k < useuse->req(); k++ ) { 1627 if( useuse->in(k) == use ) { 1628 useuse->set_req(k, r); 1629 uses_found++; 1630 } 1631 } 1632 l -= uses_found; // we deleted 1 or more copies of this edge 1633 } 1634 1635 // Now finish up 'r' 1636 r->set_req( 1, newuse ); 1637 r->set_req( 2, use ); 1638 _igvn.register_new_node_with_optimizer(r); 1639 set_loop(r, use_loop); 1640 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1641 } // End of if a loop-exit test 1642 } 1643 } 1644 1645 // Step 4: If loop-invariant use is not control, it must be dominated by a 1646 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1647 // there if needed. Make a Phi there merging old and new used values. 1648 Node_List *split_if_set = NULL; 1649 Node_List *split_bool_set = NULL; 1650 Node_List *split_cex_set = NULL; 1651 for( i = 0; i < loop->_body.size(); i++ ) { 1652 Node* old = loop->_body.at(i); 1653 Node* nnn = old_new[old->_idx]; 1654 // Copy uses to a worklist, so I can munge the def-use info 1655 // with impunity. 1656 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1657 worklist.push(old->fast_out(j)); 1658 1659 while( worklist.size() ) { 1660 Node *use = worklist.pop(); 1661 if (!has_node(use)) continue; // Ignore dead nodes 1662 if (use->in(0) == C->top()) continue; 1663 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1664 // Check for data-use outside of loop - at least one of OLD or USE 1665 // must not be a CFG node. 1666 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1667 1668 // If the Data use is an IF, that means we have an IF outside of the 1669 // loop that is switching on a condition that is set inside of the 1670 // loop. Happens if people set a loop-exit flag; then test the flag 1671 // in the loop to break the loop, then test is again outside of the 1672 // loop to determine which way the loop exited. 1673 // Loop predicate If node connects to Bool node through Opaque1 node. 1674 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1675 // Since this code is highly unlikely, we lazily build the worklist 1676 // of such Nodes to go split. 1677 if( !split_if_set ) 1678 split_if_set = new Node_List(area); 1679 split_if_set->push(use); 1680 } 1681 if( use->is_Bool() ) { 1682 if( !split_bool_set ) 1683 split_bool_set = new Node_List(area); 1684 split_bool_set->push(use); 1685 } 1686 if( use->Opcode() == Op_CreateEx ) { 1687 if( !split_cex_set ) 1688 split_cex_set = new Node_List(area); 1689 split_cex_set->push(use); 1690 } 1691 1692 1693 // Get "block" use is in 1694 uint idx = 0; 1695 while( use->in(idx) != old ) idx++; 1696 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1697 assert( !loop->is_member( get_loop( prev ) ), "" ); 1698 Node *cfg = prev->_idx >= new_counter 1699 ? prev->in(2) 1700 : idom(prev); 1701 if( use->is_Phi() ) // Phi use is in prior block 1702 cfg = prev->in(idx); // NOT in block of Phi itself 1703 if (cfg->is_top()) { // Use is dead? 1704 _igvn.replace_input_of(use, idx, C->top()); 1705 continue; 1706 } 1707 1708 while( !loop->is_member( get_loop( cfg ) ) ) { 1709 prev = cfg; 1710 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1711 } 1712 // If the use occurs after merging several exits from the loop, then 1713 // old value must have dominated all those exits. Since the same old 1714 // value was used on all those exits we did not need a Phi at this 1715 // merge point. NOW we do need a Phi here. Each loop exit value 1716 // is now merged with the peeled body exit; each exit gets its own 1717 // private Phi and those Phis need to be merged here. 1718 Node *phi; 1719 if( prev->is_Region() ) { 1720 if( idx == 0 ) { // Updating control edge? 1721 phi = prev; // Just use existing control 1722 } else { // Else need a new Phi 1723 phi = PhiNode::make( prev, old ); 1724 // Now recursively fix up the new uses of old! 1725 for( uint i = 1; i < prev->req(); i++ ) { 1726 worklist.push(phi); // Onto worklist once for each 'old' input 1727 } 1728 } 1729 } else { 1730 // Get new RegionNode merging old and new loop exits 1731 prev = old_new[prev->_idx]; 1732 assert( prev, "just made this in step 7" ); 1733 if( idx == 0 ) { // Updating control edge? 1734 phi = prev; // Just use existing control 1735 } else { // Else need a new Phi 1736 // Make a new Phi merging data values properly 1737 phi = PhiNode::make( prev, old ); 1738 phi->set_req( 1, nnn ); 1739 } 1740 } 1741 // If inserting a new Phi, check for prior hits 1742 if( idx != 0 ) { 1743 Node *hit = _igvn.hash_find_insert(phi); 1744 if( hit == NULL ) { 1745 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1746 } else { // or 1747 // Remove the new phi from the graph and use the hit 1748 _igvn.remove_dead_node(phi); 1749 phi = hit; // Use existing phi 1750 } 1751 set_ctrl(phi, prev); 1752 } 1753 // Make 'use' use the Phi instead of the old loop body exit value 1754 _igvn.replace_input_of(use, idx, phi); 1755 if( use->_idx >= new_counter ) { // If updating new phis 1756 // Not needed for correctness, but prevents a weak assert 1757 // in AddPNode from tripping (when we end up with different 1758 // base & derived Phis that will become the same after 1759 // IGVN does CSE). 1760 Node *hit = _igvn.hash_find_insert(use); 1761 if( hit ) // Go ahead and re-hash for hits. 1762 _igvn.replace_node( use, hit ); 1763 } 1764 1765 // If 'use' was in the loop-exit block, it now needs to be sunk 1766 // below the post-loop merge point. 1767 sink_use( use, prev ); 1768 } 1769 } 1770 } 1771 1772 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1773 // the loop uses a condition set in the loop. The original IF probably 1774 // takes control from one or more OLD Regions (which in turn get from NEW 1775 // Regions). In any case, there will be a set of Phis for each merge point 1776 // from the IF up to where the original BOOL def exists the loop. 1777 if( split_if_set ) { 1778 while( split_if_set->size() ) { 1779 Node *iff = split_if_set->pop(); 1780 if( iff->in(1)->is_Phi() ) { 1781 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1782 _igvn.replace_input_of(iff, 1, b); 1783 } 1784 } 1785 } 1786 if( split_bool_set ) { 1787 while( split_bool_set->size() ) { 1788 Node *b = split_bool_set->pop(); 1789 Node *phi = b->in(1); 1790 assert( phi->is_Phi(), "" ); 1791 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1792 _igvn.replace_input_of(b, 1, cmp); 1793 } 1794 } 1795 if( split_cex_set ) { 1796 while( split_cex_set->size() ) { 1797 Node *b = split_cex_set->pop(); 1798 assert( b->in(0)->is_Region(), "" ); 1799 assert( b->in(1)->is_Phi(), "" ); 1800 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1801 split_up( b, b->in(0), NULL ); 1802 } 1803 } 1804 1805 } 1806 1807 1808 //---------------------- stride_of_possible_iv ------------------------------------- 1809 // Looks for an iff/bool/comp with one operand of the compare 1810 // being a cycle involving an add and a phi, 1811 // with an optional truncation (left-shift followed by a right-shift) 1812 // of the add. Returns zero if not an iv. 1813 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1814 Node* trunc1 = NULL; 1815 Node* trunc2 = NULL; 1816 const TypeInt* ttype = NULL; 1817 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1818 return 0; 1819 } 1820 BoolNode* bl = iff->in(1)->as_Bool(); 1821 Node* cmp = bl->in(1); 1822 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1823 return 0; 1824 } 1825 // Must have an invariant operand 1826 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1827 return 0; 1828 } 1829 Node* add2 = NULL; 1830 Node* cmp1 = cmp->in(1); 1831 if (cmp1->is_Phi()) { 1832 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1833 Node* phi = cmp1; 1834 for (uint i = 1; i < phi->req(); i++) { 1835 Node* in = phi->in(i); 1836 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1837 &trunc1, &trunc2, &ttype); 1838 if (add && add->in(1) == phi) { 1839 add2 = add->in(2); 1840 break; 1841 } 1842 } 1843 } else { 1844 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1845 Node* addtrunc = cmp1; 1846 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1847 &trunc1, &trunc2, &ttype); 1848 if (add && add->in(1)->is_Phi()) { 1849 Node* phi = add->in(1); 1850 for (uint i = 1; i < phi->req(); i++) { 1851 if (phi->in(i) == addtrunc) { 1852 add2 = add->in(2); 1853 break; 1854 } 1855 } 1856 } 1857 } 1858 if (add2 != NULL) { 1859 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1860 if (add2t->is_con()) { 1861 return add2t->get_con(); 1862 } 1863 } 1864 return 0; 1865 } 1866 1867 1868 //---------------------- stay_in_loop ------------------------------------- 1869 // Return the (unique) control output node that's in the loop (if it exists.) 1870 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1871 Node* unique = NULL; 1872 if (!n) return NULL; 1873 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1874 Node* use = n->fast_out(i); 1875 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1876 if (unique != NULL) { 1877 return NULL; 1878 } 1879 unique = use; 1880 } 1881 } 1882 return unique; 1883 } 1884 1885 //------------------------------ register_node ------------------------------------- 1886 // Utility to register node "n" with PhaseIdealLoop 1887 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1888 _igvn.register_new_node_with_optimizer(n); 1889 loop->_body.push(n); 1890 if (n->is_CFG()) { 1891 set_loop(n, loop); 1892 set_idom(n, pred, ddepth); 1893 } else { 1894 set_ctrl(n, pred); 1895 } 1896 } 1897 1898 //------------------------------ proj_clone ------------------------------------- 1899 // Utility to create an if-projection 1900 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1901 ProjNode* c = p->clone()->as_Proj(); 1902 c->set_req(0, iff); 1903 return c; 1904 } 1905 1906 //------------------------------ short_circuit_if ------------------------------------- 1907 // Force the iff control output to be the live_proj 1908 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1909 guarantee(live_proj != NULL, "null projection"); 1910 int proj_con = live_proj->_con; 1911 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1912 Node *con = _igvn.intcon(proj_con); 1913 set_ctrl(con, C->root()); 1914 if (iff) { 1915 iff->set_req(1, con); 1916 } 1917 return con; 1918 } 1919 1920 //------------------------------ insert_if_before_proj ------------------------------------- 1921 // Insert a new if before an if projection (* - new node) 1922 // 1923 // before 1924 // if(test) 1925 // / \ 1926 // v v 1927 // other-proj proj (arg) 1928 // 1929 // after 1930 // if(test) 1931 // / \ 1932 // / v 1933 // | * proj-clone 1934 // v | 1935 // other-proj v 1936 // * new_if(relop(cmp[IU](left,right))) 1937 // / \ 1938 // v v 1939 // * new-proj proj 1940 // (returned) 1941 // 1942 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1943 IfNode* iff = proj->in(0)->as_If(); 1944 IdealLoopTree *loop = get_loop(proj); 1945 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1946 int ddepth = dom_depth(proj); 1947 1948 _igvn.rehash_node_delayed(iff); 1949 _igvn.rehash_node_delayed(proj); 1950 1951 proj->set_req(0, NULL); // temporary disconnect 1952 ProjNode* proj2 = proj_clone(proj, iff); 1953 register_node(proj2, loop, iff, ddepth); 1954 1955 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 1956 register_node(cmp, loop, proj2, ddepth); 1957 1958 BoolNode* bol = new BoolNode(cmp, relop); 1959 register_node(bol, loop, proj2, ddepth); 1960 1961 int opcode = iff->Opcode(); 1962 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 1963 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 1964 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 1965 register_node(new_if, loop, proj2, ddepth); 1966 1967 proj->set_req(0, new_if); // reattach 1968 set_idom(proj, new_if, ddepth); 1969 1970 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1971 guarantee(new_exit != NULL, "null exit node"); 1972 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1973 1974 return new_exit; 1975 } 1976 1977 //------------------------------ insert_region_before_proj ------------------------------------- 1978 // Insert a region before an if projection (* - new node) 1979 // 1980 // before 1981 // if(test) 1982 // / | 1983 // v | 1984 // proj v 1985 // other-proj 1986 // 1987 // after 1988 // if(test) 1989 // / | 1990 // v | 1991 // * proj-clone v 1992 // | other-proj 1993 // v 1994 // * new-region 1995 // | 1996 // v 1997 // * dum_if 1998 // / \ 1999 // v \ 2000 // * dum-proj v 2001 // proj 2002 // 2003 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2004 IfNode* iff = proj->in(0)->as_If(); 2005 IdealLoopTree *loop = get_loop(proj); 2006 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2007 int ddepth = dom_depth(proj); 2008 2009 _igvn.rehash_node_delayed(iff); 2010 _igvn.rehash_node_delayed(proj); 2011 2012 proj->set_req(0, NULL); // temporary disconnect 2013 ProjNode* proj2 = proj_clone(proj, iff); 2014 register_node(proj2, loop, iff, ddepth); 2015 2016 RegionNode* reg = new RegionNode(2); 2017 reg->set_req(1, proj2); 2018 register_node(reg, loop, iff, ddepth); 2019 2020 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2021 register_node(dum_if, loop, reg, ddepth); 2022 2023 proj->set_req(0, dum_if); // reattach 2024 set_idom(proj, dum_if, ddepth); 2025 2026 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2027 register_node(dum_proj, loop, dum_if, ddepth); 2028 2029 return reg; 2030 } 2031 2032 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2033 // Clone a signed compare loop exit from an unsigned compare and 2034 // insert it before the unsigned cmp on the stay-in-loop path. 2035 // All new nodes inserted in the dominator tree between the original 2036 // if and it's projections. The original if test is replaced with 2037 // a constant to force the stay-in-loop path. 2038 // 2039 // This is done to make sure that the original if and it's projections 2040 // still dominate the same set of control nodes, that the ctrl() relation 2041 // from data nodes to them is preserved, and that their loop nesting is 2042 // preserved. 2043 // 2044 // before 2045 // if(i <u limit) unsigned compare loop exit 2046 // / | 2047 // v v 2048 // exit-proj stay-in-loop-proj 2049 // 2050 // after 2051 // if(stay-in-loop-const) original if 2052 // / | 2053 // / v 2054 // / if(i < limit) new signed test 2055 // / / | 2056 // / / v 2057 // / / if(i <u limit) new cloned unsigned test 2058 // / / / | 2059 // v v v | 2060 // region | 2061 // | | 2062 // dum-if | 2063 // / | | 2064 // ether | | 2065 // v v 2066 // exit-proj stay-in-loop-proj 2067 // 2068 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2069 const bool Signed = true; 2070 const bool Unsigned = false; 2071 2072 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2073 if (bol->_test._test != BoolTest::lt) return NULL; 2074 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2075 if (cmpu->Opcode() != Op_CmpU) return NULL; 2076 int stride = stride_of_possible_iv(if_cmpu); 2077 if (stride == 0) return NULL; 2078 2079 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2080 guarantee(lp_proj != NULL, "null loop node"); 2081 2082 ProjNode* lp_continue = lp_proj->as_Proj(); 2083 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2084 2085 Node* limit = NULL; 2086 if (stride > 0) { 2087 limit = cmpu->in(2); 2088 } else { 2089 limit = _igvn.makecon(TypeInt::ZERO); 2090 set_ctrl(limit, C->root()); 2091 } 2092 // Create a new region on the exit path 2093 RegionNode* reg = insert_region_before_proj(lp_exit); 2094 guarantee(reg != NULL, "null region node"); 2095 2096 // Clone the if-cmpu-true-false using a signed compare 2097 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2098 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2099 reg->add_req(cmpi_exit); 2100 2101 // Clone the if-cmpu-true-false 2102 BoolTest::mask rel_u = bol->_test._test; 2103 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2104 reg->add_req(cmpu_exit); 2105 2106 // Force original if to stay in loop. 2107 short_circuit_if(if_cmpu, lp_continue); 2108 2109 return cmpi_exit->in(0)->as_If(); 2110 } 2111 2112 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2113 // Remove a previously inserted signed compare loop exit. 2114 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2115 Node* lp_proj = stay_in_loop(if_cmp, loop); 2116 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2117 stay_in_loop(lp_proj, loop)->is_If() && 2118 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2119 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2120 set_ctrl(con, C->root()); 2121 if_cmp->set_req(1, con); 2122 } 2123 2124 //------------------------------ scheduled_nodelist ------------------------------------- 2125 // Create a post order schedule of nodes that are in the 2126 // "member" set. The list is returned in "sched". 2127 // The first node in "sched" is the loop head, followed by 2128 // nodes which have no inputs in the "member" set, and then 2129 // followed by the nodes that have an immediate input dependence 2130 // on a node in "sched". 2131 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2132 2133 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2134 Arena *a = Thread::current()->resource_area(); 2135 VectorSet visited(a); 2136 Node_Stack nstack(a, loop->_body.size()); 2137 2138 Node* n = loop->_head; // top of stack is cached in "n" 2139 uint idx = 0; 2140 visited.set(n->_idx); 2141 2142 // Initially push all with no inputs from within member set 2143 for(uint i = 0; i < loop->_body.size(); i++ ) { 2144 Node *elt = loop->_body.at(i); 2145 if (member.test(elt->_idx)) { 2146 bool found = false; 2147 for (uint j = 0; j < elt->req(); j++) { 2148 Node* def = elt->in(j); 2149 if (def && member.test(def->_idx) && def != elt) { 2150 found = true; 2151 break; 2152 } 2153 } 2154 if (!found && elt != loop->_head) { 2155 nstack.push(n, idx); 2156 n = elt; 2157 assert(!visited.test(n->_idx), "not seen yet"); 2158 visited.set(n->_idx); 2159 } 2160 } 2161 } 2162 2163 // traverse out's that are in the member set 2164 while (true) { 2165 if (idx < n->outcnt()) { 2166 Node* use = n->raw_out(idx); 2167 idx++; 2168 if (!visited.test_set(use->_idx)) { 2169 if (member.test(use->_idx)) { 2170 nstack.push(n, idx); 2171 n = use; 2172 idx = 0; 2173 } 2174 } 2175 } else { 2176 // All outputs processed 2177 sched.push(n); 2178 if (nstack.is_empty()) break; 2179 n = nstack.node(); 2180 idx = nstack.index(); 2181 nstack.pop(); 2182 } 2183 } 2184 } 2185 2186 2187 //------------------------------ has_use_in_set ------------------------------------- 2188 // Has a use in the vector set 2189 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2190 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2191 Node* use = n->fast_out(j); 2192 if (vset.test(use->_idx)) { 2193 return true; 2194 } 2195 } 2196 return false; 2197 } 2198 2199 2200 //------------------------------ has_use_internal_to_set ------------------------------------- 2201 // Has use internal to the vector set (ie. not in a phi at the loop head) 2202 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2203 Node* head = loop->_head; 2204 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2205 Node* use = n->fast_out(j); 2206 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2207 return true; 2208 } 2209 } 2210 return false; 2211 } 2212 2213 2214 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2215 // clone "n" for uses that are outside of loop 2216 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2217 int cloned = 0; 2218 assert(worklist.size() == 0, "should be empty"); 2219 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2220 Node* use = n->fast_out(j); 2221 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2222 worklist.push(use); 2223 } 2224 } 2225 while( worklist.size() ) { 2226 Node *use = worklist.pop(); 2227 if (!has_node(use) || use->in(0) == C->top()) continue; 2228 uint j; 2229 for (j = 0; j < use->req(); j++) { 2230 if (use->in(j) == n) break; 2231 } 2232 assert(j < use->req(), "must be there"); 2233 2234 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2235 Node* n_clone = n->clone(); 2236 _igvn.replace_input_of(use, j, n_clone); 2237 cloned++; 2238 Node* use_c; 2239 if (!use->is_Phi()) { 2240 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2241 } else { 2242 // Use in a phi is considered a use in the associated predecessor block 2243 use_c = use->in(0)->in(j); 2244 } 2245 set_ctrl(n_clone, use_c); 2246 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2247 get_loop(use_c)->_body.push(n_clone); 2248 _igvn.register_new_node_with_optimizer(n_clone); 2249 #if !defined(PRODUCT) 2250 if (TracePartialPeeling) { 2251 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2252 } 2253 #endif 2254 } 2255 return cloned; 2256 } 2257 2258 2259 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2260 // clone "n" for special uses that are in the not_peeled region. 2261 // If these def-uses occur in separate blocks, the code generator 2262 // marks the method as not compilable. For example, if a "BoolNode" 2263 // is in a different basic block than the "IfNode" that uses it, then 2264 // the compilation is aborted in the code generator. 2265 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2266 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2267 if (n->is_Phi() || n->is_Load()) { 2268 return; 2269 } 2270 assert(worklist.size() == 0, "should be empty"); 2271 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2272 Node* use = n->fast_out(j); 2273 if ( not_peel.test(use->_idx) && 2274 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2275 use->in(1) == n) { 2276 worklist.push(use); 2277 } 2278 } 2279 if (worklist.size() > 0) { 2280 // clone "n" and insert it between inputs of "n" and the use 2281 Node* n_clone = n->clone(); 2282 loop->_body.push(n_clone); 2283 _igvn.register_new_node_with_optimizer(n_clone); 2284 set_ctrl(n_clone, get_ctrl(n)); 2285 sink_list.push(n_clone); 2286 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2287 #if !defined(PRODUCT) 2288 if (TracePartialPeeling) { 2289 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2290 } 2291 #endif 2292 while( worklist.size() ) { 2293 Node *use = worklist.pop(); 2294 _igvn.rehash_node_delayed(use); 2295 for (uint j = 1; j < use->req(); j++) { 2296 if (use->in(j) == n) { 2297 use->set_req(j, n_clone); 2298 } 2299 } 2300 } 2301 } 2302 } 2303 2304 2305 //------------------------------ insert_phi_for_loop ------------------------------------- 2306 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2307 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2308 Node *phi = PhiNode::make(lp, back_edge_val); 2309 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2310 // Use existing phi if it already exists 2311 Node *hit = _igvn.hash_find_insert(phi); 2312 if( hit == NULL ) { 2313 _igvn.register_new_node_with_optimizer(phi); 2314 set_ctrl(phi, lp); 2315 } else { 2316 // Remove the new phi from the graph and use the hit 2317 _igvn.remove_dead_node(phi); 2318 phi = hit; 2319 } 2320 _igvn.replace_input_of(use, idx, phi); 2321 } 2322 2323 #ifdef ASSERT 2324 //------------------------------ is_valid_loop_partition ------------------------------------- 2325 // Validate the loop partition sets: peel and not_peel 2326 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2327 VectorSet& not_peel ) { 2328 uint i; 2329 // Check that peel_list entries are in the peel set 2330 for (i = 0; i < peel_list.size(); i++) { 2331 if (!peel.test(peel_list.at(i)->_idx)) { 2332 return false; 2333 } 2334 } 2335 // Check at loop members are in one of peel set or not_peel set 2336 for (i = 0; i < loop->_body.size(); i++ ) { 2337 Node *def = loop->_body.at(i); 2338 uint di = def->_idx; 2339 // Check that peel set elements are in peel_list 2340 if (peel.test(di)) { 2341 if (not_peel.test(di)) { 2342 return false; 2343 } 2344 // Must be in peel_list also 2345 bool found = false; 2346 for (uint j = 0; j < peel_list.size(); j++) { 2347 if (peel_list.at(j)->_idx == di) { 2348 found = true; 2349 break; 2350 } 2351 } 2352 if (!found) { 2353 return false; 2354 } 2355 } else if (not_peel.test(di)) { 2356 if (peel.test(di)) { 2357 return false; 2358 } 2359 } else { 2360 return false; 2361 } 2362 } 2363 return true; 2364 } 2365 2366 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2367 // Ensure a use outside of loop is of the right form 2368 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2369 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2370 return (use->is_Phi() && 2371 use_c->is_Region() && use_c->req() == 3 && 2372 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2373 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2374 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2375 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2376 } 2377 2378 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2379 // Ensure that all uses outside of loop are of the right form 2380 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2381 uint orig_exit_idx, uint clone_exit_idx) { 2382 uint len = peel_list.size(); 2383 for (uint i = 0; i < len; i++) { 2384 Node *def = peel_list.at(i); 2385 2386 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2387 Node *use = def->fast_out(j); 2388 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2389 if (!loop->is_member(get_loop(use_c))) { 2390 // use is not in the loop, check for correct structure 2391 if (use->in(0) == def) { 2392 // Okay 2393 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2394 return false; 2395 } 2396 } 2397 } 2398 } 2399 return true; 2400 } 2401 #endif 2402 2403 //------------------------------ partial_peel ------------------------------------- 2404 // Partially peel (aka loop rotation) the top portion of a loop (called 2405 // the peel section below) by cloning it and placing one copy just before 2406 // the new loop head and the other copy at the bottom of the new loop. 2407 // 2408 // before after where it came from 2409 // 2410 // stmt1 stmt1 2411 // loop: stmt2 clone 2412 // stmt2 if condA goto exitA clone 2413 // if condA goto exitA new_loop: new 2414 // stmt3 stmt3 clone 2415 // if !condB goto loop if condB goto exitB clone 2416 // exitB: stmt2 orig 2417 // stmt4 if !condA goto new_loop orig 2418 // exitA: goto exitA 2419 // exitB: 2420 // stmt4 2421 // exitA: 2422 // 2423 // Step 1: find the cut point: an exit test on probable 2424 // induction variable. 2425 // Step 2: schedule (with cloning) operations in the peel 2426 // section that can be executed after the cut into 2427 // the section that is not peeled. This may need 2428 // to clone operations into exit blocks. For 2429 // instance, a reference to A[i] in the not-peel 2430 // section and a reference to B[i] in an exit block 2431 // may cause a left-shift of i by 2 to be placed 2432 // in the peel block. This step will clone the left 2433 // shift into the exit block and sink the left shift 2434 // from the peel to the not-peel section. 2435 // Step 3: clone the loop, retarget the control, and insert 2436 // phis for values that are live across the new loop 2437 // head. This is very dependent on the graph structure 2438 // from clone_loop. It creates region nodes for 2439 // exit control and associated phi nodes for values 2440 // flow out of the loop through that exit. The region 2441 // node is dominated by the clone's control projection. 2442 // So the clone's peel section is placed before the 2443 // new loop head, and the clone's not-peel section is 2444 // forms the top part of the new loop. The original 2445 // peel section forms the tail of the new loop. 2446 // Step 4: update the dominator tree and recompute the 2447 // dominator depth. 2448 // 2449 // orig 2450 // 2451 // stmt1 2452 // | 2453 // v 2454 // loop predicate 2455 // | 2456 // v 2457 // loop<----+ 2458 // | | 2459 // stmt2 | 2460 // | | 2461 // v | 2462 // ifA | 2463 // / | | 2464 // v v | 2465 // false true ^ <-- last_peel 2466 // / | | 2467 // / ===|==cut | 2468 // / stmt3 | <-- first_not_peel 2469 // / | | 2470 // | v | 2471 // v ifB | 2472 // exitA: / \ | 2473 // / \ | 2474 // v v | 2475 // false true | 2476 // / \ | 2477 // / ----+ 2478 // | 2479 // v 2480 // exitB: 2481 // stmt4 2482 // 2483 // 2484 // after clone loop 2485 // 2486 // stmt1 2487 // | 2488 // v 2489 // loop predicate 2490 // / \ 2491 // clone / \ orig 2492 // / \ 2493 // / \ 2494 // v v 2495 // +---->loop loop<----+ 2496 // | | | | 2497 // | stmt2 stmt2 | 2498 // | | | | 2499 // | v v | 2500 // | ifA ifA | 2501 // | | \ / | | 2502 // | v v v v | 2503 // ^ true false false true ^ <-- last_peel 2504 // | | ^ \ / | | 2505 // | cut==|== \ \ / ===|==cut | 2506 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2507 // | | dom | | | | 2508 // | v \ 1v v2 v | 2509 // | ifB regionA ifB | 2510 // | / \ | / \ | 2511 // | / \ v / \ | 2512 // | v v exitA: v v | 2513 // | true false false true | 2514 // | / ^ \ / \ | 2515 // +---- \ \ / ----+ 2516 // dom \ / 2517 // \ 1v v2 2518 // regionB 2519 // | 2520 // v 2521 // exitB: 2522 // stmt4 2523 // 2524 // 2525 // after partial peel 2526 // 2527 // stmt1 2528 // | 2529 // v 2530 // loop predicate 2531 // / 2532 // clone / orig 2533 // / TOP 2534 // / \ 2535 // v v 2536 // TOP->loop loop----+ 2537 // | | | 2538 // stmt2 stmt2 | 2539 // | | | 2540 // v v | 2541 // ifA ifA | 2542 // | \ / | | 2543 // v v v v | 2544 // true false false true | <-- last_peel 2545 // | ^ \ / +------|---+ 2546 // +->newloop \ \ / === ==cut | | 2547 // | stmt3 \ \ / TOP | | 2548 // | | dom | | stmt3 | | <-- first_not_peel 2549 // | v \ 1v v2 v | | 2550 // | ifB regionA ifB ^ v 2551 // | / \ | / \ | | 2552 // | / \ v / \ | | 2553 // | v v exitA: v v | | 2554 // | true false false true | | 2555 // | / ^ \ / \ | | 2556 // | | \ \ / v | | 2557 // | | dom \ / TOP | | 2558 // | | \ 1v v2 | | 2559 // ^ v regionB | | 2560 // | | | | | 2561 // | | v ^ v 2562 // | | exitB: | | 2563 // | | stmt4 | | 2564 // | +------------>-----------------+ | 2565 // | | 2566 // +-----------------<---------------------+ 2567 // 2568 // 2569 // final graph 2570 // 2571 // stmt1 2572 // | 2573 // v 2574 // loop predicate 2575 // | 2576 // v 2577 // stmt2 clone 2578 // | 2579 // v 2580 // ........> ifA clone 2581 // : / | 2582 // dom / | 2583 // : v v 2584 // : false true 2585 // : | | 2586 // : | v 2587 // : | newloop<-----+ 2588 // : | | | 2589 // : | stmt3 clone | 2590 // : | | | 2591 // : | v | 2592 // : | ifB | 2593 // : | / \ | 2594 // : | v v | 2595 // : | false true | 2596 // : | | | | 2597 // : | v stmt2 | 2598 // : | exitB: | | 2599 // : | stmt4 v | 2600 // : | ifA orig | 2601 // : | / \ | 2602 // : | / \ | 2603 // : | v v | 2604 // : | false true | 2605 // : | / \ | 2606 // : v v -----+ 2607 // RegionA 2608 // | 2609 // v 2610 // exitA 2611 // 2612 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2613 2614 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2615 if (!loop->_head->is_Loop()) { 2616 return false; } 2617 2618 LoopNode *head = loop->_head->as_Loop(); 2619 2620 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2621 return false; 2622 } 2623 2624 // Check for complex exit control 2625 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2626 Node *n = loop->_body.at(ii); 2627 int opc = n->Opcode(); 2628 if (n->is_Call() || 2629 opc == Op_Catch || 2630 opc == Op_CatchProj || 2631 opc == Op_Jump || 2632 opc == Op_JumpProj) { 2633 #if !defined(PRODUCT) 2634 if (TracePartialPeeling) { 2635 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2636 } 2637 #endif 2638 return false; 2639 } 2640 } 2641 2642 int dd = dom_depth(head); 2643 2644 // Step 1: find cut point 2645 2646 // Walk up dominators to loop head looking for first loop exit 2647 // which is executed on every path thru loop. 2648 IfNode *peel_if = NULL; 2649 IfNode *peel_if_cmpu = NULL; 2650 2651 Node *iff = loop->tail(); 2652 while( iff != head ) { 2653 if( iff->is_If() ) { 2654 Node *ctrl = get_ctrl(iff->in(1)); 2655 if (ctrl->is_top()) return false; // Dead test on live IF. 2656 // If loop-varying exit-test, check for induction variable 2657 if( loop->is_member(get_loop(ctrl)) && 2658 loop->is_loop_exit(iff) && 2659 is_possible_iv_test(iff)) { 2660 Node* cmp = iff->in(1)->in(1); 2661 if (cmp->Opcode() == Op_CmpI) { 2662 peel_if = iff->as_If(); 2663 } else { 2664 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2665 peel_if_cmpu = iff->as_If(); 2666 } 2667 } 2668 } 2669 iff = idom(iff); 2670 } 2671 // Prefer signed compare over unsigned compare. 2672 IfNode* new_peel_if = NULL; 2673 if (peel_if == NULL) { 2674 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2675 return false; // No peel point found 2676 } 2677 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2678 if (new_peel_if == NULL) { 2679 return false; // No peel point found 2680 } 2681 peel_if = new_peel_if; 2682 } 2683 Node* last_peel = stay_in_loop(peel_if, loop); 2684 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2685 if (first_not_peeled == NULL || first_not_peeled == head) { 2686 return false; 2687 } 2688 2689 #if !defined(PRODUCT) 2690 if (TraceLoopOpts) { 2691 tty->print("PartialPeel "); 2692 loop->dump_head(); 2693 } 2694 2695 if (TracePartialPeeling) { 2696 tty->print_cr("before partial peel one iteration"); 2697 Node_List wl; 2698 Node* t = head->in(2); 2699 while (true) { 2700 wl.push(t); 2701 if (t == head) break; 2702 t = idom(t); 2703 } 2704 while (wl.size() > 0) { 2705 Node* tt = wl.pop(); 2706 tt->dump(); 2707 if (tt == last_peel) tty->print_cr("-- cut --"); 2708 } 2709 } 2710 #endif 2711 ResourceArea *area = Thread::current()->resource_area(); 2712 VectorSet peel(area); 2713 VectorSet not_peel(area); 2714 Node_List peel_list(area); 2715 Node_List worklist(area); 2716 Node_List sink_list(area); 2717 2718 // Set of cfg nodes to peel are those that are executable from 2719 // the head through last_peel. 2720 assert(worklist.size() == 0, "should be empty"); 2721 worklist.push(head); 2722 peel.set(head->_idx); 2723 while (worklist.size() > 0) { 2724 Node *n = worklist.pop(); 2725 if (n != last_peel) { 2726 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2727 Node* use = n->fast_out(j); 2728 if (use->is_CFG() && 2729 loop->is_member(get_loop(use)) && 2730 !peel.test_set(use->_idx)) { 2731 worklist.push(use); 2732 } 2733 } 2734 } 2735 } 2736 2737 // Set of non-cfg nodes to peel are those that are control 2738 // dependent on the cfg nodes. 2739 uint i; 2740 for(i = 0; i < loop->_body.size(); i++ ) { 2741 Node *n = loop->_body.at(i); 2742 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2743 if (peel.test(n_c->_idx)) { 2744 peel.set(n->_idx); 2745 } else { 2746 not_peel.set(n->_idx); 2747 } 2748 } 2749 2750 // Step 2: move operations from the peeled section down into the 2751 // not-peeled section 2752 2753 // Get a post order schedule of nodes in the peel region 2754 // Result in right-most operand. 2755 scheduled_nodelist(loop, peel, peel_list ); 2756 2757 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2758 2759 // For future check for too many new phis 2760 uint old_phi_cnt = 0; 2761 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2762 Node* use = head->fast_out(j); 2763 if (use->is_Phi()) old_phi_cnt++; 2764 } 2765 2766 #if !defined(PRODUCT) 2767 if (TracePartialPeeling) { 2768 tty->print_cr("\npeeled list"); 2769 } 2770 #endif 2771 2772 // Evacuate nodes in peel region into the not_peeled region if possible 2773 uint new_phi_cnt = 0; 2774 uint cloned_for_outside_use = 0; 2775 for (i = 0; i < peel_list.size();) { 2776 Node* n = peel_list.at(i); 2777 #if !defined(PRODUCT) 2778 if (TracePartialPeeling) n->dump(); 2779 #endif 2780 bool incr = true; 2781 if ( !n->is_CFG() ) { 2782 2783 if ( has_use_in_set(n, not_peel) ) { 2784 2785 // If not used internal to the peeled region, 2786 // move "n" from peeled to not_peeled region. 2787 2788 if ( !has_use_internal_to_set(n, peel, loop) ) { 2789 2790 // if not pinned and not a load (which maybe anti-dependent on a store) 2791 // and not a CMove (Matcher expects only bool->cmove). 2792 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2793 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2794 sink_list.push(n); 2795 peel >>= n->_idx; // delete n from peel set. 2796 not_peel <<= n->_idx; // add n to not_peel set. 2797 peel_list.remove(i); 2798 incr = false; 2799 #if !defined(PRODUCT) 2800 if (TracePartialPeeling) { 2801 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2802 n->_idx, get_ctrl(n)->_idx); 2803 } 2804 #endif 2805 } 2806 } else { 2807 // Otherwise check for special def-use cases that span 2808 // the peel/not_peel boundary such as bool->if 2809 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2810 new_phi_cnt++; 2811 } 2812 } 2813 } 2814 if (incr) i++; 2815 } 2816 2817 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2818 #if !defined(PRODUCT) 2819 if (TracePartialPeeling) { 2820 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2821 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2822 } 2823 #endif 2824 if (new_peel_if != NULL) { 2825 remove_cmpi_loop_exit(new_peel_if, loop); 2826 } 2827 // Inhibit more partial peeling on this loop 2828 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2829 head->mark_partial_peel_failed(); 2830 if (cloned_for_outside_use > 0) { 2831 // Terminate this round of loop opts because 2832 // the graph outside this loop was changed. 2833 C->set_major_progress(); 2834 return true; 2835 } 2836 return false; 2837 } 2838 2839 // Step 3: clone loop, retarget control, and insert new phis 2840 2841 // Create new loop head for new phis and to hang 2842 // the nodes being moved (sinked) from the peel region. 2843 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2844 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2845 _igvn.register_new_node_with_optimizer(new_head); 2846 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2847 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2848 set_loop(new_head, loop); 2849 loop->_body.push(new_head); 2850 not_peel.set(new_head->_idx); 2851 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2852 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2853 2854 while (sink_list.size() > 0) { 2855 Node* n = sink_list.pop(); 2856 set_ctrl(n, new_head); 2857 } 2858 2859 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2860 2861 clone_loop( loop, old_new, dd ); 2862 2863 const uint clone_exit_idx = 1; 2864 const uint orig_exit_idx = 2; 2865 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2866 2867 Node* head_clone = old_new[head->_idx]; 2868 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2869 Node* orig_tail_clone = head_clone->in(2); 2870 2871 // Add phi if "def" node is in peel set and "use" is not 2872 2873 for(i = 0; i < peel_list.size(); i++ ) { 2874 Node *def = peel_list.at(i); 2875 if (!def->is_CFG()) { 2876 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2877 Node *use = def->fast_out(j); 2878 if (has_node(use) && use->in(0) != C->top() && 2879 (!peel.test(use->_idx) || 2880 (use->is_Phi() && use->in(0) == head)) ) { 2881 worklist.push(use); 2882 } 2883 } 2884 while( worklist.size() ) { 2885 Node *use = worklist.pop(); 2886 for (uint j = 1; j < use->req(); j++) { 2887 Node* n = use->in(j); 2888 if (n == def) { 2889 2890 // "def" is in peel set, "use" is not in peel set 2891 // or "use" is in the entry boundary (a phi) of the peel set 2892 2893 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2894 2895 if ( loop->is_member(get_loop( use_c )) ) { 2896 // use is in loop 2897 if (old_new[use->_idx] != NULL) { // null for dead code 2898 Node* use_clone = old_new[use->_idx]; 2899 _igvn.replace_input_of(use, j, C->top()); 2900 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2901 } 2902 } else { 2903 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2904 // use is not in the loop, check if the live range includes the cut 2905 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2906 if (not_peel.test(lp_if->_idx)) { 2907 assert(j == orig_exit_idx, "use from original loop"); 2908 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2909 } 2910 } 2911 } 2912 } 2913 } 2914 } 2915 } 2916 2917 // Step 3b: retarget control 2918 2919 // Redirect control to the new loop head if a cloned node in 2920 // the not_peeled region has control that points into the peeled region. 2921 // This necessary because the cloned peeled region will be outside 2922 // the loop. 2923 // from to 2924 // cloned-peeled <---+ 2925 // new_head_clone: | <--+ 2926 // cloned-not_peeled in(0) in(0) 2927 // orig-peeled 2928 2929 for(i = 0; i < loop->_body.size(); i++ ) { 2930 Node *n = loop->_body.at(i); 2931 if (!n->is_CFG() && n->in(0) != NULL && 2932 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2933 Node* n_clone = old_new[n->_idx]; 2934 _igvn.replace_input_of(n_clone, 0, new_head_clone); 2935 } 2936 } 2937 2938 // Backedge of the surviving new_head (the clone) is original last_peel 2939 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 2940 2941 // Cut first node in original not_peel set 2942 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 2943 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 2944 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 2945 2946 // Copy head_clone back-branch info to original head 2947 // and remove original head's loop entry and 2948 // clone head's back-branch 2949 _igvn.rehash_node_delayed(head); // Multiple edge updates 2950 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2951 head->set_req(LoopNode::LoopBackControl, C->top()); 2952 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 2953 2954 // Similarly modify the phis 2955 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2956 Node* use = head->fast_out(k); 2957 if (use->is_Phi() && use->outcnt() > 0) { 2958 Node* use_clone = old_new[use->_idx]; 2959 _igvn.rehash_node_delayed(use); // Multiple edge updates 2960 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2961 use->set_req(LoopNode::LoopBackControl, C->top()); 2962 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 2963 } 2964 } 2965 2966 // Step 4: update dominator tree and dominator depth 2967 2968 set_idom(head, orig_tail_clone, dd); 2969 recompute_dom_depth(); 2970 2971 // Inhibit more partial peeling on this loop 2972 new_head_clone->set_partial_peel_loop(); 2973 C->set_major_progress(); 2974 loop->record_for_igvn(); 2975 2976 #if !defined(PRODUCT) 2977 if (TracePartialPeeling) { 2978 tty->print_cr("\nafter partial peel one iteration"); 2979 Node_List wl(area); 2980 Node* t = last_peel; 2981 while (true) { 2982 wl.push(t); 2983 if (t == head_clone) break; 2984 t = idom(t); 2985 } 2986 while (wl.size() > 0) { 2987 Node* tt = wl.pop(); 2988 if (tt == head) tty->print_cr("orig head"); 2989 else if (tt == new_head_clone) tty->print_cr("new head"); 2990 else if (tt == head_clone) tty->print_cr("clone head"); 2991 tt->dump(); 2992 } 2993 } 2994 #endif 2995 return true; 2996 } 2997 2998 //------------------------------reorg_offsets---------------------------------- 2999 // Reorganize offset computations to lower register pressure. Mostly 3000 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3001 // then alive with the post-incremented trip counter forcing an extra 3002 // register move) 3003 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3004 // Perform it only for canonical counted loops. 3005 // Loop's shape could be messed up by iteration_split_impl. 3006 if (!loop->_head->is_CountedLoop()) 3007 return; 3008 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3009 return; 3010 3011 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3012 CountedLoopEndNode *cle = cl->loopexit(); 3013 Node *exit = cle->proj_out(false); 3014 Node *phi = cl->phi(); 3015 3016 // Check for the special case of folks using the pre-incremented 3017 // trip-counter on the fall-out path (forces the pre-incremented 3018 // and post-incremented trip counter to be live at the same time). 3019 // Fix this by adjusting to use the post-increment trip counter. 3020 3021 bool progress = true; 3022 while (progress) { 3023 progress = false; 3024 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3025 Node* use = phi->fast_out(i); // User of trip-counter 3026 if (!has_ctrl(use)) continue; 3027 Node *u_ctrl = get_ctrl(use); 3028 if (use->is_Phi()) { 3029 u_ctrl = NULL; 3030 for (uint j = 1; j < use->req(); j++) 3031 if (use->in(j) == phi) 3032 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3033 } 3034 IdealLoopTree *u_loop = get_loop(u_ctrl); 3035 // Look for loop-invariant use 3036 if (u_loop == loop) continue; 3037 if (loop->is_member(u_loop)) continue; 3038 // Check that use is live out the bottom. Assuming the trip-counter 3039 // update is right at the bottom, uses of of the loop middle are ok. 3040 if (dom_lca(exit, u_ctrl) != exit) continue; 3041 // Hit! Refactor use to use the post-incremented tripcounter. 3042 // Compute a post-increment tripcounter. 3043 Node *opaq = new Opaque2Node( C, cle->incr() ); 3044 register_new_node(opaq, exit); 3045 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3046 set_ctrl(neg_stride, C->root()); 3047 Node *post = new AddINode( opaq, neg_stride); 3048 register_new_node(post, exit); 3049 _igvn.rehash_node_delayed(use); 3050 for (uint j = 1; j < use->req(); j++) { 3051 if (use->in(j) == phi) 3052 use->set_req(j, post); 3053 } 3054 // Since DU info changed, rerun loop 3055 progress = true; 3056 break; 3057 } 3058 } 3059 3060 }